等价标准形:
A ∈ C n × n , ∃ 可 逆 P ∈ C m × m , Q ∈ C n × n A \in C^{n \times n},\exist 可逆P \in C^{m \times m},\ Q \in C^{n \times n} A∈Cn×n,∃可逆P∈Cm×m, Q∈Cn×n,使 A = P [ I r 0 0 0 ] Q A = P \begin{bmatrix} I_r & 0 \\ 0 & 0 \end{bmatrix} Q A=P[Ir000]Q
相似标准形:
A ∈ C n × n , ∃ 可 逆 P ∈ C n × n A \in C^{n \times n},\exist 可逆P \in C^{n \times n} A∈Cn×n,∃可逆P∈Cn×n,使 A = P [ λ 1 λ 2 ⋱ λ n ] P − 1 A = P \begin{bmatrix} \lambda_1 & & \\ & \lambda_2 & \\ & & \ddots & \\ & & & \lambda_n \end{bmatrix} P^{-1} A=P⎣⎢⎢⎡λ1λ2⋱λn⎦⎥⎥⎤P−1或 A = P J A P − 1 A = P J_A P^{-1} A=PJAP−1
A ∈ R n × n , A T = A , 则 ∃ 正 交 P ( P T P = I ) A \in R^{n \times n},\ A^T = A,则\exist 正交P(P^T P = I) A∈Rn×n, AT=A,则∃正交P(PTP=I),使 P T A P = C − 1 A C = d i a g ( λ 1 , λ 2 , ⋯ , λ n ) P^T A P = C^{-1} A C = diag(\lambda_1,\ \lambda_2,\ \cdots,\ \lambda_n) PTAP=C−1AC=diag(λ1, λ2, ⋯, λn)
设 U = [ a 11 a 12 ⋯ a 1 n a 21 ⋯ a 2 n ⋱ ⋮ a n n ] U = \begin{bmatrix} a_{11} & a_{12} & \cdots & a_{1n} \\ & a_{21} & \cdots & a_{2n} \\ & & \ddots & \vdots \\ & & & a_{nn} \end{bmatrix} U=⎣⎢⎢⎢⎡a11a12a21⋯⋯⋱a1na2n⋮ann⎦⎥⎥⎥⎤为上三角矩阵, L = [ a 11 a 21 a 22 ⋮ ⋮ ⋱ a n 1 a n 2 ⋯ a n n ] L = \begin{bmatrix} a_{11} & & & \\ a_{21} & a_{22} & & \\ \vdots& \vdots & \ddots & \\ a_{n1} & a_{n2} & \cdots & a_{nn} \end{bmatrix} L=⎣⎢⎢⎢⎡a11a21⋮an1a22⋮an2⋱⋯ann⎦⎥⎥⎥⎤为下三角矩阵
LU分解: A = L U A = LU A=LU
LDV分解: A = L D V A = LDV A=LDV,其中V的对角线元素全为1,D是对角矩阵
下三角可逆矩阵P,使 P A = U ⇒ A = P − 1 U PA=U \Rightarrow A = P^{-1} U PA=U⇒A=P−1U
A的三角分解方法:
1. 对A作初等行变换, ( A , I ) = ( U , P ) (A,\ I) = (U,\ P) (A, I)=(U, P)
2. L = P − 1 L = P^{-1} L=P−1, ( P , I ) = ( I , P − 1 ) (P,\ I) = (I,\ P^{-1}) (P, I)=(I, P−1)
3. U = DV,其中D=diag(U), V是对U的对角线元素归一化后的矩阵
例如:
A有唯一LDV分解
⟺ \iff ⟺
A的顺序主子式 Δ k = ∣ a 11 ⋯ a 1 k a 21 ⋯ a 2 k ⋮ ⋮ a k 1 ⋯ a k k ∣ ≠ 0 , k = 1 , 2 , ⋯ , n − 1 ; Δ 0 = 1 \Delta_k = \begin{vmatrix} a_{11} & \cdots & a_{1k} \\ a_{21} & \cdots & a_{2k} \\ \vdots & & \vdots \\ a_{k1} & \cdots & a_{kk} \end{vmatrix} \neq 0,k = 1,\ 2,\ \cdots,\ n - 1;\Delta_0 = 1 Δk=∣∣∣∣∣∣∣∣∣a11a21⋮ak1⋯⋯⋯a1ka2k⋮akk∣∣∣∣∣∣∣∣∣=0,k=1, 2, ⋯, n−1;Δ0=1
其中 D = [ d 1 d 2 ⋱ d n ] , d k = Δ k Δ k − 1 ; k = 1 , 2 , ⋯ , n D = \begin{bmatrix} d_1 & & & \\ & d_2 & & \\ & & \ddots & \\ & & & d_n \end{bmatrix},d_k = \frac{\Delta_k}{\Delta_{k - 1}};k = 1,\ 2,\ \cdots,\ n D=⎣⎢⎢⎡d1d2⋱dn⎦⎥⎥⎤,dk=Δk−1Δk;k=1, 2, ⋯, n
A = ( a i j ) ∈ F n × n , r a n k ( A ) = k ( k ≤ n ) , A 的 顺 序 主 子 式 Δ j ≠ 0 , j = 1 , 2 , ⋯ , k A = (a_{ij}) \in F^{n \times n},\ rank(A) = k(k \leq n),\ A的顺序主子式\Delta_j \neq 0, j = 1,\ 2,\ \cdots,\ k A=(aij)∈Fn×n, rank(A)=k(k≤n), A的顺序主子式Δj=0,j=1, 2, ⋯, k,则A有LU分解
不是所有的矩阵都有LU分解,例如 A = [ 0 1 1 0 ] A = \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix} A=[0110]
可逆矩阵 A ∈ F n × n A \in F^{n \times n} A∈Fn×n,则A有LU分解 ⟺ \iff ⟺A的所有顺序主子式 Δ k ≠ 0 , k = 1 , 2 , ⋯ , n − 1 \Delta_k \neq 0,\ k = 1,\ 2,\ \cdots,\ n - 1 Δk=0, k=1, 2, ⋯, n−1
例如:
设A的LU分解为 A = L U A = LU A=LU,则 A X = b ⟺ L U X = b ⟺ { L Y = b U X = Y AX = b \iff LU X = b \iff \begin{cases} LY = b \\ UX = Y \end{cases} AX=b⟺LUX=b⟺{LY=bUX=Y
满秩分解: A ∈ F m × n , r a n k ( A ) = r , 若 ∃ 秩 为 r 的 矩 阵 B ∈ F m × r , C ∈ F r × n , 使 A = B C A \in F^{m \times n},\ rank(A) = r,若\exist秩为r的矩阵B \in F^{m \times r},\ C \in F^{r \times n},使A = BC A∈Fm×n, rank(A)=r,若∃秩为r的矩阵B∈Fm×r, C∈Fr×n,使A=BC
任何非零矩阵 A ∈ F m × n A \in F^{m \times n} A∈Fm×n,都存在满秩分解。A的满秩分解一般不惟一
求满秩分解方法:
方法3
Hermite标准形:
例如:
A = [ 1 1 2 0 2 2 1 0 1 ] A = \begin{bmatrix} 1 & 1 & 2 \\ 0 & 2 & 2 \\ 1 & 0 & 1 \end{bmatrix} A=⎣⎡101120221⎦⎤,求A的满秩分解
解:
A = [ 1 1 2 0 2 2 1 0 1 ] → [ 1 0 1 0 1 1 0 0 0 ] ⇒ r a n k ( A ) = 2 A = \begin{bmatrix} 1 & 1 & 2 \\ 0 & 2 & 2 \\ 1 & 0 & 1 \end{bmatrix} \rightarrow \begin{bmatrix} 1 & 0 & 1 \\ 0 & 1 & 1 \\ 0 & 0 & 0 \end{bmatrix} \Rightarrow rank(A) = 2 A=⎣⎡101120221⎦⎤→⎣⎡100010110⎦⎤⇒rank(A)=2
则 B = [ 1 1 0 2 1 0 ] , C = [ 1 0 1 0 1 1 ] , A = B C B = \begin{bmatrix} 1 & 1 \\ 0 & 2 \\ 1 & 0 \end{bmatrix},\quad C = \begin{bmatrix} 1 & 0 & 1 \\ 0 & 1 & 1 \end{bmatrix},\ A = BC B=⎣⎡101120⎦⎤,C=[100111], A=BC
谱: 方阵A的所有互异特征值得集合
谱分解: A = ∑ i = 1 s λ i P i A = \sum \limits^s_{i = 1}\lambda_i P_i A=i=1∑sλiPi,即可对角化矩阵可分解为s个方阵 P i P_i Pi的加权和
分解过程如下:
幂等矩阵: P 2 = P P^2 = P P2=P
性质:
P H \quad P^H PH和 ( I − P ) (I - P) (I−P)仍是幂等矩阵
\quad P的特征值为1或者是0,而且P可相似于对角矩阵
F n = N ( P ) ⨁ R ( P ) \quad F^n = N(P) \bigoplus R(P) Fn=N(P)⨁R(P)
A ∈ C n × n A \in C^{n \times n} A∈Cn×n,A可对角化 ⟺ A = ∑ i = 1 s λ i P i \iff A = \sum^{s}_{i = 1} \lambda_i P_i ⟺A=∑i=1sλiPi
其中 λ i \lambda_i λi为A的谱, P i ∈ C n × n P_i \in C^{n \times n} Pi∈Cn×n 满足
1. P i 2 = P i , i = 1 , 2 , ⋯ , s \quad 1.\ P_i^2 = P_i, i = 1,\ 2,\ \cdots,\ s 1. Pi2=Pi,i=1, 2, ⋯, s
2. P i P j = 0 , i ≠ j \quad 2.\ P_i P_j = 0,\ i \neq j 2. PiPj=0, i=j
3. ∑ i = 1 s P i = I n \quad 3.\ \sum \limits^s_{i = 1}P_i = I_n 3. i=1∑sPi=In
半正定Hermite矩阵: A ∈ F n × n , A H = A , ∀ x ∈ F n , 有 x H A X ≥ 0 A \in F^{n \times n},\ A^H = A,\ \forall x \in F^n,有x^H A X \geq 0 A∈Fn×n, AH=A, ∀x∈Fn,有xHAX≥0,则称A为半正定的。即
A为半正定的 ⟺ \iff ⟺ A的特征值为非负实数
A ∈ F n × n A \in F^{n \times n} A∈Fn×n是半正定的Hermite矩阵, r a n k ( A ) = k rank(A) = k rank(A)=k,则 A = v 1 v 1 H + v 2 v 2 H + ⋯ + v k v k H A = v_1 v_1^H + v_2 v_2^H + \cdots + v_k v_k^H A=v1v1H+v2v2H+⋯+vkvkH
v i ∈ F n , { v 1 , v 2 , ⋯ , v k } v_i \in F^n,\ \{v_1,\ v_2,\ \cdots,\ v_k\} vi∈Fn, {v1, v2, ⋯, vk}是空间 F n F^n Fn中非零的正交向量组
正交(酉)相似: A , B ∈ R n × n , ∃ 正 交 ( 酉 ) 矩 阵 U , 使 U T A U = B ( U H A U = B ) A,\ B \in R^{n \times n},\ \exist 正交(酉)矩阵U,使U^T A U = B(U^H A U = B) A, B∈Rn×n, ∃正交(酉)矩阵U,使UTAU=B(UHAU=B),则称A正交(酉)相似于B
UR分解: A ∈ C n × n , ∣ A ∣ ≠ 0 A \in C^{n \times n},\ |A| \neq 0 A∈Cn×n, ∣A∣=0,则存在酉矩阵 U ∈ C n × n U \in C^{n \times n} U∈Cn×n及主对角线上全为正的上三角阵
R = [ r 11 r 12 ⋯ r 1 n r 22 ⋯ r 2 n ⋱ ⋮ r n n ] , r i i > 0 R = \begin{bmatrix} r_{11} & r_{12} & \cdots & r_{1n} \\ & r_{22} & \cdots & r_{2n} \\ & & \ddots & \vdots \\ & & & r_{nn} \end{bmatrix}, \quad r_{ii} > 0 R=⎣⎢⎢⎢⎡r11r12r22⋯⋯⋱r1nr2n⋮rnn⎦⎥⎥⎥⎤,rii>0,使 A = U R A = UR A=UR
UR分解求解步骤:
1. 取A的列向量记为 A 1 , ⋯ A n A_1, \cdots A_n A1,⋯An
2. 对 A 1 , ⋯ A n A_1, \cdots A_n A1,⋯An使用gram-schmidt正交化,得到标准正交向量组U
3. 根据 A = U R A = UR A=UR求出R,使用初等行变换: ( U ∣ A ) → ( E ∣ R ) (U|A) \rightarrow (E|R) (U∣A)→(E∣R)
例如:
QR分解:
A ∈ C m × k A \in C^{m \times k} A∈Cm×k是一个列满秩(列向量组线性无关)的矩阵,即 r a n k ( A ) = k rank(A) = k rank(A)=k,则A可被分解为 A = Q R A = QR A=QR。
其中Q为列规范矩阵(列向量两两正交且为单位向量),R是可逆的上三角矩阵
QR分解求解步骤:
1. 取A的列向量记为 A 1 , ⋯ A n A_1, \cdots A_n A1,⋯An
2. 对 A 1 , ⋯ A n A_1, \cdots A_n A1,⋯An使用gram-schmidt正交化,得到标准正交向量组Q
3. 根据 A = Q R A = QR A=QR求出R,使用初等行变换: ( Q ∣ A ) → ( E ∣ R ) (Q|A) \rightarrow (E|R) (Q∣A)→(E∣R)
Schur分解: A ∈ C n × n A \in C^{n \times n} A∈Cn×n,存在酉矩阵U和上三角矩阵T,使得 U H A U = T = [ λ 1 t 12 ⋯ t 1 n λ 2 ⋯ t 2 n ⋱ ⋮ λ n ] U^H A U = T = \begin{bmatrix} \lambda_1 & t_{12} & \cdots & t_{1n} \\ & \lambda_2 & \cdots & t_{2n} \\ & & \ddots & \vdots \\ & & & \lambda_n \end{bmatrix} UHAU=T=⎣⎢⎢⎢⎡λ1t12λ2⋯⋯⋱t1nt2n⋮λn⎦⎥⎥⎥⎤
其中 λ i \lambda_i λi为矩阵A的特征值, i = 1 , 2 , ⋯ , n i = 1,\ 2,\ \cdots,\ n i=1, 2, ⋯, n
Schur分解求解步骤:
1. 当不存在Jordan矩阵 J A J_A JA和可逆矩阵P,使 J A = P − 1 A P J_A = P^{-1} A P JA=P−1AP时,需要求Jordan矩阵 J A J_A JA和可逆矩阵P。求解步骤见Jordan标准形计算步骤
2. 取P的列向量记为 ( α 1 , ⋯ α n ) (\alpha_1, \cdots \alpha_n) (α1,⋯αn)
3. 对 ( α 1 , ⋯ α n ) (\alpha_1, \cdots \alpha_n) (α1,⋯αn)使用gram-schmidt正交化,得到标准正交向量组U
4. 根据 P = U R P = UR P=UR求出R,使用初等行变换: ( U ∣ P ) → ( E ∣ R ) (U|P) \rightarrow (E|R) (U∣P)→(E∣R)
5. 令 T = R J A R − 1 T = R J_A R^{-1} T=RJAR−1,即得 U H A U = T U^H A U = T UHAU=T
正规矩阵:
A ∈ R n × n , A T A = A A T ; A ∈ C n × n , A H A = A A H A \in R^{n \times n},A^T A= A A^T; \quad A \in C^{n \times n},A^H A= A A^H A∈Rn×n,ATA=AAT;A∈Cn×n,AHA=AAH
对角矩阵、Hermite矩阵、反Hermite矩阵( A H = − A A^H = -A AH=−A)、酉(正交)矩阵、实对称矩阵均为正规矩阵
复对称矩阵不是正规的,例如 A = [ 1 i i − 1 ] , A H = [ 1 − i − i − 1 ] , 但 A H A ≠ A A H A = \begin{bmatrix}1 & i \\ i & -1 \end{bmatrix},\ A^H = \begin{bmatrix}1 & -i \\ -i & -1 \end{bmatrix},\ 但A^H A \neq A A^H A=[1ii−1], AH=[1−i−i−1], 但AHA=AAH
例如:
A ∈ C n × n A \in C^{n \times n} A∈Cn×n是正规矩阵 ⟺ \iff ⟺ A酉相似于对角矩阵,即 ∃ U ∈ C n × n , U H A U = d i a g ( λ 1 , λ 2 , ⋯ , λ n ) \exist U \in C^{n \times n},\ U^H A U = diag(\lambda_1, \lambda_2,\ \cdots,\ \lambda_n) ∃U∈Cn×n, UHAU=diag(λ1,λ2, ⋯, λn)
A ∈ C n × n A \in C^{n \times n} A∈Cn×n是正规矩阵 ⟺ \iff ⟺A有n个线性无关的特征向量构成空间 C n C^n Cn的标准正交基
A ∈ F n × n A \in F^{n \times n} A∈Fn×n,若A是正规矩阵,则A必有n个线性无关的特征向量,且不同特征值对应的特征向量线性无关
正规矩阵谱分解求解步骤:
1. 通过 A H A = A A H A^H A = A A^H AHA=AAH判断A是否是正规矩阵
2. 求A的特征值, ∣ λ I − A ∣ = ( λ − λ 1 ) ⋯ ( λ − λ n ) = 0 |\lambda I - A| = (\lambda - \lambda_1) \cdots (\lambda - \lambda_n) = 0 ∣λI−A∣=(λ−λ1)⋯(λ−λn)=0
3. 求A的对应的线性无关的特征向量: ϵ 1 , ⋯ , ϵ n \epsilon_1, \cdots,\ \epsilon_n ϵ1,⋯, ϵn
4. 将 ϵ 1 , ⋯ , ϵ n \epsilon_1, \cdots,\ \epsilon_n ϵ1,⋯, ϵn标准正交化,得到 u 1 , ⋯ , u n u_1,\ \cdots,\ u_n u1, ⋯, un
5. 令 U = ( u 1 , ⋯ , u n ) U = (u_1,\ \cdots,\ u_n) U=(u1, ⋯, un),即得 U H A U = d i a g ( λ 1 , ⋯ , λ n ) U^H A U = diag(\lambda_1,\ \cdots,\ \lambda_n) UHAU=diag(λ1, ⋯, λn)
正交矩阵的谱分解: A ∈ C n × n , λ ( A ) = { λ 1 , ⋯ , λ n } A \in C^{n \times n},\ \lambda(A) = \{\lambda_1,\ \cdots,\ \lambda_n\} A∈Cn×n, λ(A)={λ1, ⋯, λn},则
A是正规矩阵 ⟺ \iff ⟺ A有如下谱分解, A = ∑ i = 1 s λ i P i , p i ∈ C n × n A = \sum \limits^s_{i = 1} \lambda_i P_i,\ p_i \in C^{n \times n} A=i=1∑sλiPi, pi∈Cn×n,
其中 P i 2 = P i , P i H = P i , i = 1 , ⋯ , n P i P J = 0 , i ≠ j I n = ∑ i = 1 s P i \quad P^2_i = P_i,\ P^H_i = P_i,\ i = 1, \cdots,\ n \\ \qquad \ \ P_i P_J = 0,\ i \neq j \\ \qquad \ \ I_n = \sum \limits^s_{i = 1} P_i Pi2=Pi, PiH=Pi, i=1,⋯, n PiPJ=0, i=j In=i=1∑sPi
例如:
A = [ 0 − 1 i 1 0 0 i 0 0 ] A = \begin{bmatrix} 0 & -1 & i \\ 1 & 0 & 0 \\ i & 0 & 0 \end{bmatrix} A=⎣⎡01i−100i00⎦⎤,求酉矩阵U和A的谱分解,使 U H A U = ∧ U^H A U = \land UHAU=∧
解:
酉矩阵U: A H A = A A H ∣ λ I − A ∣ = λ ( λ 2 + 2 ) = 0 ⇒ λ 1 = 2 i , λ 2 = − 2 i , λ 3 = 0 对 应 的 特 征 向 量 ϵ 1 = ( 2 , − i , 1 ) T , ϵ 2 = ( − 2 , − i , 1 ) T , ϵ 3 = ( 0 , − 1 , i ) T 将 ϵ 1 , ϵ 2 , ϵ 3 正 交 化 , u i = ϵ i ∣ ∣ ϵ i ∣ ∣ 得 u 1 = ( 2 2 , − i 2 , 1 2 ) T , u 2 = ( − 2 2 , − i 2 , 1 2 ) T , u 3 = ( 0 , − 1 2 , − i 2 ) 令 U = ( u 1 , u 2 , u 3 ) , 则 U 为 酉 矩 阵 , 且 U H A U = d i a g ( 2 i , − 2 i , 0 ) A^H A = A A^H \\ |\lambda_I - A| = \lambda(\lambda^2+2) = 0 \Rightarrow \lambda_1 = \sqrt{2} i,\ \lambda_2 = -\sqrt{2} i,\ \lambda_3 = 0 \\ 对应的特征向量\epsilon_1 = (\sqrt{2},\ -i,\ 1)^T,\ \epsilon_2 = (-\sqrt{2},\ -i,\ 1)^T,\ \epsilon_3 = (0,\ -1,\ i)^T \\ 将\epsilon_1,\epsilon_2,\epsilon_3正交化,u_i = \frac{\epsilon_i}{||\epsilon_i||}得 \\ u_1 = (\frac{\sqrt{2}}2,\ -\frac{i}{2},\ \frac{1}{2})^T,u_2 = (-\frac{\sqrt{2}}2,\ -\frac{i}{2},\ \frac{1}{2})^T,u_3 = (0,\ -\frac{1}{\sqrt{2}},\ -\frac{i}{\sqrt{2}}) \\ 令U = (u_1,\ u_2,\ u_3),则U为酉矩阵,且U^H A U = diag(\sqrt{2} i,\ -\sqrt{2} i,\ 0) AHA=AAH∣λI−A∣=λ(λ2+2)=0⇒λ1=2i, λ2=−2i, λ3=0对应的特征向量ϵ1=(2, −i, 1)T, ϵ2=(−2, −i, 1)T, ϵ3=(0, −1, i)T将ϵ1,ϵ2,ϵ3正交化,ui=∣∣ϵi∣∣ϵi得u1=(22, −2i, 21)T,u2=(−22, −2i, 21)T,u3=(0, −21, −2i)令U=(u1, u2, u3),则U为酉矩阵,且UHAU=diag(2i, −2i, 0)
\newline
A的谱分解: A = ( u 1 , u 2 , u 3 ) [ 2 i ( 1 0 0 ) − 2 i ( 0 1 0 ) + 0 ( 0 0 1 ) ] ( u 1 H u 2 H u 3 H ) = 2 i u 1 u 1 H − 2 i u 2 u 2 H + 0 u 3 u 3 H \begin{aligned}A & = (u_1,\ u_2,\ u_3) \left[ \sqrt{2} i \begin{pmatrix} 1 & & \\ & 0 & \\ & & 0 \end{pmatrix} -\sqrt{2} i \begin{pmatrix} 0 & & \\ & 1 & \\ & & 0 \end{pmatrix} +0\begin{pmatrix} 0 & & \\ & 0 & \\ & & 1 \end{pmatrix} \right] \begin{pmatrix} u_1^H \\ u_2^H \\ u_3^H \end{pmatrix} \\ & = \sqrt{2}i u_1 u^H_1 - \sqrt{2}i u_2 u_2^H + 0u_3 u_3^H \end{aligned} A=(u1, u2, u3)⎣⎡2i⎝⎛100⎠⎞−2i⎝⎛010⎠⎞+0⎝⎛001⎠⎞⎦⎤⎝⎛u1Hu2Hu3H⎠⎞=2iu1u1H−2iu2u2H+0u3u3H
∀ A ∈ C m × n \forall A \in C^{m \times n} ∀A∈Cm×n,有 A H A ∈ C n × n , A A H ∈ C m × m A^HA \in C^{n \times n},AA^H \in C^{m \times m} AHA∈Cn×n,AAH∈Cm×m,且 A H A , A A H A^HA,AA^H AHA,AAH均为Hermite矩阵
设 A ∈ C m × n A \in C^{m \times n} A∈Cm×n,则
秩(A) = 秩( A H A A^HA AHA) = 秩( A A H AA^H AAH)
A H A 与 A A H A^HA与AA^H AHA与AAH的非零特征值相等,其特征值为非负实数
A H A 与 A A H A^HA与AA^H AHA与AAH均为半正定矩阵,当秩(A)=n时, A H A A^HA AHA为正定的
奇异值: 设 A ∈ C m × n A \in C^{m \times n} A∈Cm×n,秩(A) = r > 0,矩阵 A H A A^HA AHA的特征值为
λ 1 ≥ λ 2 ≥ ⋯ ≥ λ r > 0 , λ r + 1 = ⋯ = λ n = 0 ( \lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_r > 0,\ \lambda_{r+1} = \cdots = \lambda_n = 0( λ1≥λ2≥⋯≥λr>0, λr+1=⋯=λn=0(矩阵 A H A A^HA AHA的特征值为 λ 1 ≥ λ 2 ≥ ⋯ λ r > 0 , λ r + 1 = ⋯ = λ n = 0 ) \lambda_1 \ge \lambda_2 \ge \cdots \lambda_r > 0,\ \lambda_{r+1} = \cdots = \lambda_n = 0) λ1≥λ2≥⋯λr>0, λr+1=⋯=λn=0)
称正数 σ i = λ i ( i = 1 , ⋯ , r ) \sigma_i = \sqrt{\lambda_i}(i = 1,\ \cdots,\ r) σi=λi(i=1, ⋯, r)为A的奇异值
奇异值的求法:
A ∈ C m × n A \in C^{m \times n} A∈Cm×n,求 A H A A^HA AHA的特征值: λ 1 ≥ ⋯ ≥ λ r > 0 \lambda_1 \ge \cdots \ge \lambda_r > 0 λ1≥⋯≥λr>0,即得 σ i = λ i \sigma_i = \sqrt{\lambda_i} σi=λi
A ∈ C n × n A \in C^{n \times n} A∈Cn×n为正规矩阵,则A的奇异值为A的非零特征值的模,即 σ = ∣ λ ∣ , λ ∈ λ ( A ) \sigma = |\lambda|, \lambda \in \lambda(A) σ=∣λ∣,λ∈λ(A)
A ∈ C n × n A \in C^{n \times n} A∈Cn×n为正定的Hermite矩阵时,A的奇异值为A的特征值
酉等价: ∃ B ∈ C m × n , 酉 矩 阵 U ∈ C m × m , V ∈ C n × n , 使 U A V = B \exists B \in C^{m \times n},酉矩阵U \in C^{m \times m}, V \in C^{n \times n},使UAV = B ∃B∈Cm×n,酉矩阵U∈Cm×m,V∈Cn×n,使UAV=B
酉等价的矩阵A,B具有相同的奇异值。
矩阵的谱分解是奇异分解的特例,因为谱分解的前提是正规矩阵,然后奇异分解针对的是一般情况。
∀ A ∈ C m × n , 秩 ( A ) = r > 0 , σ 1 ≥ ⋯ ≥ σ r \forall A \in C^{m \times n},秩(A) = r > 0, \sigma_1 \ge \cdots \ge \sigma_r ∀A∈Cm×n,秩(A)=r>0,σ1≥⋯≥σr为A的奇异值,则存在酉矩阵 U ∈ C m × m , V ∈ n × n U \in C^{m \times m},\ V \in {n \times n} U∈Cm×m, V∈n×n,使
A m × n = U m × m ( Δ r 0 ) m × n V n × n H , 其 中 Δ r = ( σ 1 ⋱ σ r ) A_{m \times n} = U_{m \times m} \begin{pmatrix} \Delta_r & \\ & 0 \end{pmatrix}_{m \times n} V^H_{n \times n},其中\Delta_r = \begin{pmatrix}\sigma_1 & & \\ & \ddots & \\ & & \sigma_r\end{pmatrix} Am×n=Um×m(Δr0)m×nVn×nH,其中Δr=⎝⎛σ1⋱σr⎠⎞
A的奇异值分解不唯一
A ∈ C m × n , U = ( u 1 , ⋯ , u m ) ∈ C m × m , V = ( v 1 , ⋯ , v n ) ∈ C n × n A \in C^{m \times n},\ U = (u_1,\ \cdots,\ u_m) \in C^{m \times m}, V=(v_1,\ \cdots,\ v_n) \in C^{n \times n} A∈Cm×n, U=(u1, ⋯, um)∈Cm×m,V=(v1, ⋯, vn)∈Cn×n,则
A = U Σ V H = ( u 1 , ⋯ , u m ) ( σ 1 ⋱ σ r 0 ) ( v 1 H ⋮ v n H ) = σ 1 u 1 v 1 H + ⋯ + σ r u r v r H A = U \Sigma V^H = (u_1,\ \cdots,\ u_m) \begin{pmatrix}\sigma_1 & & & \\ & \ddots & & \\ & & \sigma_r & \\ & & & 0\end{pmatrix} \begin{pmatrix}v_1^H \\ \vdots \\ v_n^H\end{pmatrix} = \sigma_1 u_1 v_1^H + \cdots + \sigma_r u_r v_r^H A=UΣVH=(u1, ⋯, um)⎝⎜⎜⎛σ1⋱σr0⎠⎟⎟⎞⎝⎜⎛v1H⋮vnH⎠⎟⎞=σ1u1v1H+⋯+σrurvrH
左奇异向量: U = { u 1 , u 2 , ⋯ , u r } U=\{u_1,\ u_2,\ \cdots,\ u_r\} U={u1, u2, ⋯, ur}
右奇异向量: V = { v 1 , v 2 , ⋯ , v r } V=\{v_1,\ v_2,\ \cdots,\ v_r\} V={v1, v2, ⋯, vr}
A ∈ C m × n A \in C^{m \times n} A∈Cm×n,A的奇异值 σ 1 ≥ ⋯ ≥ σ r > 0 \sigma_1 \ge \cdots \ge \sigma_r > 0 σ1≥⋯≥σr>0,则 A = σ 1 u 1 v 1 H + ⋯ + σ r u r v r H A= \sigma_1 u_1 v_1^H + \cdots + \sigma_r u_r v_r^H A=σ1u1v1H+⋯+σrurvrH
A H A = A A H ⟺ Σ i = 1 n ∣ λ i ∣ 2 ≤ t r ( A H A ) = t r ( A A H ) = Σ i = 1 n σ i 2 A^HA = AA^H \iff \Sigma^n_{i = 1} |\lambda_i|^2 \leq tr(A^HA) = tr(AA^H) = \Sigma^n_{i = 1} \sigma_i ^2 AHA=AAH⟺Σi=1n∣λi∣2≤tr(AHA)=tr(AAH)=Σi=1nσi2
求A的SVD步骤:
1. 求 A H A A^HA AHA的特征值: σ 1 2 ≥ ⋯ ≥ σ r 2 , σ r + 1 2 = ⋯ = σ n 2 = 0 \sigma_1\ ^2 \ge \cdots \ge \sigma_r\ ^2,\sigma_{r + 1}\ ^2 = \cdots = \sigma_n\ ^2 = 0 σ1 2≥⋯≥σr 2,σr+1 2=⋯=σn 2=0
A H A A^HA AHA对应线性无关的特征向量: ϵ 1 , ⋯ , ϵ r , ⋯ , ϵ n \epsilon_1, \cdots, \epsilon_r, \cdots, \epsilon_n ϵ1,⋯,ϵr,⋯,ϵn
2. 将 ϵ 1 , ⋯ , ϵ r , ⋯ , ϵ n \epsilon_1, \cdots, \epsilon_r, \cdots, \epsilon_n ϵ1,⋯,ϵr,⋯,ϵn标准正交化: v 1 , ⋯ , v n v_1, \cdots, v_n v1,⋯,vn,取 V = ( v 1 , ⋯ , v n ) V = (v_1, \cdots, v_n) V=(v1,⋯,vn)
3. 令 u i = 1 σ i A v i , i = 1 , ⋯ , r u_i = \frac{1}{\sigma_i}Av_i,i = 1, \cdots, r ui=σi1Avi,i=1,⋯,r,将 u 1 , ⋯ , u r u_1,\ \cdots, u_r u1, ⋯,ur扩充为 c m c^m cm的标准正交基 u 1 , ⋯ , u m u_1, \cdots, u_m u1,⋯,um,令 U = ( u 1 , ⋯ , u m ) U = (u_1, \cdots, u_m) U=(u1,⋯,um)即得
A = U ( σ 1 ⋱ σ r 0 ) V H A = U \begin{pmatrix}\sigma_1 & & & \\ & \ddots & & \\ & & \sigma_r & \\ & & & 0\end{pmatrix}V^H A=U⎝⎜⎜⎛σ1⋱σr0⎠⎟⎟⎞VH
注:
A H A = V ( Δ r 2 0 ) n × n V H A^HA = V \begin{pmatrix}\Delta_r\ ^2 & \\ & 0\end{pmatrix}_{n \times n} V^H AHA=V(Δr 20)n×nVH
A A H = U ( Δ r 2 0 ) m × m U H AA^H = U \begin{pmatrix}\Delta_r\ ^2 & \\ & 0\end{pmatrix}_{m \times m} U^H AAH=U(Δr 20)m×mUH